> Getting back to the original problem, I'm wondering if a more detailed > explanation of what is trying to be measured and why might help. Is it > crypto performance (in which case, the crypto apis might be useful to > look at), or something else? The root of the issue was my attempt to compare integer performance on a T1000 with other machines that I am working on. This was the result of rather poor performance I got on the T1000 running a web server with dynamic workload, compared with a fairly recent 4-core Xeon. The libgcrypt-based benchmark was supposed to extract the maximum out of the T1000 (which indeed it did, as all 32 strands reported close to 250 million instructions per second). It then occurred to me that the user-space environment was 32 bits, and I was wondering if 64 bit code would perhaps work better.
Just for reference, here are the results of this benchmark on different machines. The results are specified in times to encrypt a 1 GB file, with the number of parallel threads set to the number of execution units for each machine: A (2 x 3.06 GHz Xeon model 2 stepping 5 with HT enabled): 1m29s (4 threads) B (4 x 2.80 GHz Xeon model 2 stepping 5): 1m07s (4 threads) C (1 GHz UltraSPARC T1, 8 cores, 32 strands): 48s (32 threads) D (2.83 GHz Xeon E5440 4 cores): 21s (4 threads) --Elad _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
